Synthetic GPR B-Scan Augmentation for Noise-Robust AI Training
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
The challenge of training AI models to recognize targets in ground penetrating radar (GPR) B-scan data is exacerbated by numerous variables affecting radar signals, including target size, shape, orientation, material, distance, soil composition, moisture content, conductivity, frequency content, signal strength, antenna design, data collection mode, and signal processing, necessitating a large and costly real data set.
Innovation Solution
The method involves generating synthetic data to mimic observed blurring, RF interference, vertical shifts, and low-pass filtering in GPR B-scan images, using techniques such as Gaussian blurring, RF noise simulation, and high-pass filtering to create a training data set that accounts for these variables.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Reliability
If real GPR data is collected to train AI models, then the training data set accurately represents real-world conditions, but the time and cost to collect such data becomes substantial and prohibitive
Solution Approach 1:
The patent creates synthetic GPR B-scan images that copy and mimic the characteristics of real GPR data. These synthetic images replicate blurring effects, RF interference patterns, vertical shifts, and other artifacts found in real radar data, allowing AI models to be trained on realistic data without the time and cost constraints of collecting actual field data
Solution Approach 2:
The patent applies multiple image processing operations that change parameters of the synthetic images to simulate various real-world conditions. By adjusting blur kernels, adding noise with specific statistical properties, applying vertical shifts, and modifying other image characteristics, the system generates diverse training data that represents different soil conditions, target configurations, and radar operating parameters
2Reliability
If real GPR data is collected to train AI models, then the training data set accurately represents real-world conditions, but the cost to collect such data becomes substantial and prohibitive
Solution Approach 1:
Instead of expending resources to collect real GPR data across diverse conditions, the patent uses computational methods to generate synthetic images that copy the essential characteristics of real data. This approach eliminates the need for expensive field campaigns, equipment deployment, and data collection operations while producing training data that maintains realism through careful replication of physical artifacts
Solution Approach 2:
The patent replaces the mechanical process of physical data collection with digital image processing operations. Rather than deploying radar equipment in the field and manually acquiring data, the system uses software-based simulations and image manipulation to generate training data, substituting computational mechanisms for physical measurement systems
3Productivity
If synthetic data is generated to reduce time and cost, then the training data set can be created more efficiently, but identifying all significant variables and simulating their variation becomes non-trivial
Solution Approach 1:
The patent breaks down the complex task of simulating all possible GPR variations into discrete, manageable image processing operations. Each operation addresses a specific artifact or characteristic (blurring for speed variations, noise addition for RF interference, vertical shifts for ground movement), making the overall process more systematic and easier to implement than attempting to simulate all variables simultaneously
Solution Approach 2:
The patent performs preliminary actions by first obtaining real GPR images and extracting their characteristic artifacts, then uses these extracted features to guide the synthetic data generation process. This preliminary analysis of real data ensures that the synthetic generation process targets the most significant variables and uses empirically observed patterns rather than making assumptions about which factors are most important
Data Source
AI summary
Methods of making a partially synthetic data set for training an AI to correct for noise in GPR B-scan data. The methods mimic noise observed in real B-scans. Original B-scans are augmented with a plurality of permutations of synthetic noise. Methods include processes for (i) mimicking observed blurring in B-scan images resulting from variations in translational speed of the detector, (ii) mimicking observed RF interference in B-scan images, and/or (iii) mimicking observed vertical shifts resulting from movement of the GPR unit due to uneven ground. Methods can further include processes for low-pass filtering B-scan images to remove inductive emission interference and artifacts resulting from dynamic range limitations.